AI

GPT-6 Astra Drops a Block in a Bowl 19 Times in 20. The Puzzle Still Stalls

Robb Harlan 4 min read

RoboCurve posted hardware numbers for GPT-6 Astra on 4 September. Same agent policy, same two tasks, dual I2RT YAM arms: put a red block in a bowl, and seat a round puzzle piece by its knob into a matching groove.

The bowl is the headline. Astra finished 19 of 20. Anthropic’s Claude Fable 5.1 was 8 of 20. Fable 5 was 1 of 20. Mean time 2.5 minutes vs 6.8 and 8.2. Estimated list-price cost $0.94 per run vs $2.12 and $2.69. Output tokens 2.1k vs 12.9k and 19.2k.

The puzzle is the hangover. Astra 2 of 20. Fable 5.1 2 of 20. Fable 5 0 of 20. Astra still gets there faster and cheaper (3.4 min, $1.36), then dies at the last press.

Cost-vs-success scatter next to overhead and wrist views of YAM arms moving a red letter block toward a white bowl
Bowl-task comparison still from RoboCurve’s Astra vs Fable 5.1 video. Source: RoboCurve.

How the bench is wired

The RoboCurve write-up is specific:

  • Embodiment: bimanual I2RT YAM, 6-DoF per arm, parallel-jaw grippers
  • Command: absolute end-effector poses (move_to) plus gripper; onboard IK
  • Observation: top camera, left wrist, right wrist, proprioception
  • Policy: Inspect Robots agent, medium thinking, 20 LLM-call budget, 25% speed cap
  • Scoring: human grader, stages 0–4 (no approach through seated)

Astra’s mean stage on the bowl is 3.95. On the puzzle it is 2.00. The write-up says it consistently reaches the groove (stage 3) and then hesitates or misaligns the press.

Humanoids Daily repeats those counts and notes Inspect Robots as the open harness (GitHub: robocurve/inspect-robots).

Scatter of completion rate versus API cost beside robot-arm video feeds of the block-in-bowl task
Humanoids Daily’s composite of the RoboCurve bowl comparison. Source: Humanoids Daily / RoboCurve.

Caveats they printed themselves

RoboCurve lists the holes. Puzzle trials shared rig-4. Bowl trials did not: Fable ran on rig-3, Astra on rig-1 because rig-3 was down. Astra’s runs were two days later, not interleaved. Graders knew the model. Costs are list price; OpenAI cached about a fifth of Astra’s input, so Astra’s dollar figure is, if anything, high.

That is enough to keep me from treating 95% as a factory number. It is still a clean gap on the easy task and a clean tie on the hard one.

A Human’s Take

Gross pick-and-place is becoming a language-model parlor trick. The groove is not. If your policy cannot feel the last millimeter, it will look brilliant on a bowl and helpless on a peg. I would run this bench again interleaved, on one rig, with a tactile channel in the loop. Until then, Astra is a fast, cheap kinematic clerk. It is not a fitter.

Sources